[clang] [clang-tools-extra] [clang][NFC] Refactor `CXXNewExpr::InitializationStyle` (re-land) (PR #71417)

[via cfe-commits]

tomasz-kaminski-sonarsource wrote:

> What I don't want to lose from this patch are the changes to places like `InitializationStyle getInitializationStyle() const` and `CXXNewExpr::CXXNewExpr` where the old code was unclear and the new code is significantly more clear. We should not be performing math on the enumerator values to encode in-band extra information. What I see being clarified by this patch is:
> 
> ```
> struct S { int x; };
> auto *s0 = new int; // None, scalar types have no notional constructor initialization
> auto *s1 = new S; // Implicit, class type has a notional constructor call
> auto *s2 = new S(0); // Call (ParenList is a much better name)
> auto *s3 = new S{0}; // List (BraceList is a much better name)
> ```
> 
> > In both cases, the news would report the initialization as Implicit, where actually no initialization is performed. There is no call to the constructor inserted.
> 
> There is an implicit constructor call but it's a noop because the type is trivial, so I think `Implicit` is what I would expect given the comment `/// New-expression has no written initializer, but has an implicit one.` https://godbolt.org/z/353G45vnc
> 
> That said, I can see why it may be confusing to say there's an implicit initialization for something that is a noop which performs no initialization and we have an enumerator for "no initialization". I think @tomasz-kaminski-sonarsource would like for the extra enumerator to be removed, but I don't think that's possible to do without also losing the benefits of the changes. But perhaps we could rename `Implicit` and `None` to something more clear?

Yes, this is the thing that I am aiming for.
As far as changes LLVM code goes, I see that everywhere expect `hasInitializer()` where `NoInit` is used, then `Implicit` is handled in the same manner, so as far as this particular PR goes removing the `Implicit` will make the code simpler. I do not think that changing public-facing API to eliminate storage-hack is the right direction. 

> 
> > In short, as the downstream that uses AST for writing rules, we will need to update all the uses of NoInit to also check for Implicit, without getting any value from the distinction.
> 
> The C++ APIs have no stability guarantees and not every change will be to the benefit of all downstreams; I see the changes in this PR as being an improvement over the status quo because they clarify code in our code base and I'm not seeing the same level of confusion you are in your downstream. (That said, I'm also totally happy to rename enumeration members to pick more descriptive names.)

With this change, I am no longer able to explain what the meaning of this enumeration is supposed to be. Previously it was all about the syntax, and the meaning of values was clear. Now we separate two cases, where we differentiate lack of initializer being spelled between doing no constructor calls and doing possibly implicit constructor calls. 
If we go with that direction, we should also differentiate between parents (`Call`) doing various semantics effects, such as calling constructor, construction aggregate, or initializing built-in types. Same for the braces. This would lead to an unwieldy number of the enumerator's values, which illustrates the problem of trying to shoehorn two orthogonal aspects into one enumerator.
For example, if we decide to introduce AST that would represent performing default initialization of trivial object (for example to capture Erroneous Behavior), all uses of `NoInit` would turn into `Implicit` because there was an initializer node.

I am less concerned about the number of changes required downstream, but the long-standing impact on the API. Where I would need to explain that the `NoInit` value does not mean that:
 * no initializer was written  (because it is also Implicit)
 * no initialization was performed (because `Implicit` also means that).
 

https://github.com/llvm/llvm-project/pull/71417